亚马逊和英伟达的关系刚刚变得更加紧密。两家公司周三宣布扩大合作伙伴关系,其中包括一项协议,将在亚马逊的数据中心新增200万块英伟达GPU芯片。
这些GPU专为满足训练和运行AI模型的重度计算需求而设计,包括英伟达Blackwell Ultra、Rubin和Rubin Ultra GPU。这些芯片将于2027年和2028年进入亚马逊网络服务(AWS)的数据中心。
这一消息是在英伟达季度财报电话会议期间宣布的,距离亚马逊同意从今年开始在AWS基础设施中部署超过100万块英伟达GPU仅过去了五个月。英伟达在一份声明中表示,自那以来,“需求已超出预期”。
两家公司均未披露财务条款。目前尚不清楚英伟达的确切回报是多少。但考虑到GPU的单价成本,这笔交易的规模高达数百亿美元。
这一公告之所以引人注目,不仅在于其规模之大和增长之快,还因为它超出了亚马逊单纯采购更多英伟达芯片的范畴。而且,这一合作是在亚马逊同时投资自研、可能形成竞争的AI芯片的背景下达成的。
英伟达周三表示,其技术——包括将数千块GPU连接成一个系统的网络硬件,以及其开源模型、CPU、数据处理软件和机器人平台——也将全面集成到AWS中。
两家公司表示,来自初创公司、企业、AI实验室乃至政府的“需求激增”影响了双方更紧密合作的决定。
此次扩大合作之际,亚马逊正在加大自研AI芯片的力度——尤其是在CPU方面,CPU是服务器核心的通用处理器。
亚马逊一直在自研芯片,以减少对英伟达的依赖,甚至与这家芯片巨头展开竞争。亚马逊AI负责人Peter DeSantis曾表示,AWS正在洽谈将其Trainium芯片——在深度学习工作负载方面是英伟达H100或Blackwell芯片的直接替代品——出售给其他公司用于数据中心。亚马逊基于Arm架构的Graviton CPU也被视为英特尔和AMD传统服务器芯片的挑战者。
亚马逊表示其定制芯片业务正在增长,在最近一次财报电话会议上指出,该业务年化收入运行率已突破 250 亿美元,这得益于 Anthropic 和 OpenAI 等 AI 实验室总计 2250 亿美元的承诺投入。
但看起来,Nvidia 仍然是 AI 芯片领域的王者。
随着亚马逊从第三季度开始向 AWS 新增 200 万颗 GPU 芯片,Nvidia 还计划发送数量不详的 Vera CPU,“其中一些与 Rubin 集成,另一些则独立出货”,据 Nvidia 首席财务官 Colette Kress 表示。
Nvidia 首席执行官黄仁勋对公司的 Vera CPU 抱有宏大计划,早在 5 月他就夸口称,为公司找到了一个“全新的 2000 亿美元 TAM(潜在市场)”。
除了 AWS 之外,Kress 周三表示,Nvidia 预计 Vera 将被“所有主流超大规模云厂商、新型云服务商、AI 实验室和系统 OEM 厂商部署,且向主要合作伙伴的出货已经在进行中”,这些合作伙伴包括 Oracle 和 SpaceX AI。
这一合作还延伸到了亚马逊的仓储机器人和企业级产品。
Kress 表示,亚马逊计划采用 Nvidia 的完整物理 AI 技术栈来驱动其机器人队伍。该技术栈包括 Omniverse(其仿真与数字孪生平台)、Cosmos(其世界模型平台)、Isaac(其机器人开发平台)以及 Jetson(面向机器人和边缘 AI 的计算硬件)。本周,Nvidia 还推出了新版本的 Jetson,定位为一款更易用的机器人计算机,面向“入门级边缘 AI”。
在企业端,AWS 将在 Amazon Bedrock(其托管基础模型平台)和 SageMaker(其托管云服务)上提供 Nvidia 的 Nemotron 系列开源模型。
Nvidia 周三还报告称,第二季度销售额达到 962 亿美元,超出分析师预期。数据中心收入占 Nvidia 本季度销售额的大部分,达到 890 亿美元,同比增长 117%。
Nvidia 表示,预计第三季度营收将达到 1080 亿美元,其中一部分将来自其下一代 Rubin GPU。Nvidia 表示,本季度已开始生产出货。投资者一直在关注 Rubin 在第三季度的首批销售情况,以判断对其下一代硬件的需求是否会持续。
英伟达已承诺投入 2790 亿美元,用于保障当前及未来数据中心项目的供应和制造产能,较上季度的 1190 亿美元大幅增加。这家芯片制造商正寻求锁定内存和制造产能,以满足未来几年的人工智能需求。该承诺包括本财年剩余时间预计支出的 920 亿美元,以及 2028 财年另外的 870 亿美元。
黄仁勋在周三的电话会议上表示:“对行业而言,重要的是人工智能现在正在从事有生产力、有价值的工作。AI 正在生成可盈利的 token……如果我们有更多算力,就能生成更多可盈利的 token,这将为所有服务带来更多利润。这正是我们目前所处的阶段,也是所有人都在加码投入的原因。”
随着 AI 公司向基础设施投入数千亿美元,投资者将关注额外算力是否确实能如此直接地转化为额外利润。
Amazon and Nvidia just got a lot closer. The two companies announced Wednesday an expanded partnership that includes a deal to add another 2 million Nvidia GPU chips to Amazon’s data centers.
These GPUs, which are designed to handle the heavy compute demands of training and running AI models, include Nvidia Blackwell Ultra, Rubin, and Rubin Ultra GPUs. The chips will head to Amazon Web Services’ data centers in 2027 and 2028.
The announcement, made during Nvidia’s quarterly earnings call, comes just five months after Amazon agreed to deploy more than 1 million Nvidia GPUs across AWS infrastructure starting this year. Nvidia said in a statement that since then, “demand has exceeded those expectations.”
Neither company shared financial terms. It’s unclear what the exact return will be for Nvidia. But considering GPU units costs, the deal is worth tens of billions of dollars.
The announcement is notable not just for its size and the speed in which it grew, but also because it extends beyond Amazon buying more Nvidia chips. And it’s happening even as Amazon invests in its own potentially competing AI chips.
Nvidia said Wednesday that its technology, including the networking hardware that connects thousands of GPUs into one system, as well as its open models, CPUs, data processing software, and robotics platform, will also be integrated across AWS.
The companies said “surging demand” from startups, enterprises, AI labs, and even governments influenced the decision to work more closely.
The expanded partnership comes as Amazon ramps up its own AI chip efforts — particularly with CPUs, which are the general purpose processors at the heart of servers.
Amazon has been building its own chips to lessen its dependence on Nvidia and even compete with the chip giant. Amazon’s AI chief Peter DeSantis has said that AWS is in talks to sell its Trainium chips — which are a direct alternative to Nvidia’s H100 or Blackwell chips for deep learning workloads — to other companies for use in data centers. Amazon’s Arm-built Graviton CPU is also seen as a challenger to traditional server chips from Intel and AMD.
Amazon has said its custom chip business is growing, noting on its last earnings call that it crossed a $25 billion annualized revenue run rate, driven by $225 billion in total commitments from AI labs like Anthropic and OpenAI.
But, it seems Nvidia is still the GOAT in the world of AI chips.
With the 2 million GPU chips Amazon is adding to AWS starting in the third quarter, Nvidia also plans to send an unspecified number of Vera CPUs, “some integrated with Rubin, others standalone,” according to Nvidia CFO Colette Kress.
Nvidia CEO Jensen Huang has big plans for the company’s Vera CPUs, boasting back in May that he had found a “brand new $200 billion TAM” for the company.
Aside from AWS, Kress said Wednesday that Nvidia expects Vera to be deployed by “every major hyperscaler, neocloud, AI lab, and system OEM, with shipments already underway to our lead partners,” which include Oracle and SpaceX AI.
The partnership is also extending to Amazon’s warehouse robots and enterprise offerings.
Kress said Amazon plans to adopt Nvidia’s full physical AI stack to power its fleet of robots. The stack includes Omniverse (its simulation and digital twin platform); Cosmos (its world model platform); Isaac (its robotics development platform); and Jetson (computing hardware for robots and edge AI). This week, Nvidia also introduced a new version of Jetson designed as a more accessible robotics computer for “entry-level edge AI.”
On the enterprise side, AWS will serve Nvidia’s Nemotron family of open models on Amazon Bedrock, its managed foundation model platform, and SageMaker, its managed cloud service.
Nvidia also reported Wednesday that it recorded sales of $96.2 billion for the second quarter, beating analyst estimates. Data center revenue made up the majority of Nvidia’s sales for the quarter at $89 billion, up 117% from a year ago.
Nvidia said it expects revenue to reach $108 billion in the third quarter, some of which will come from its next-gen Rubin GPUs. Nvidia said it began production shipments this quarter. Investors have been looking out for Rubin’s initial Q3 sales for signs that demand will continue into Nvidia’s next generation of hardware.
Nvidia has committed $279 billion to secure supply and manufacturing capacity for current and future data-center projects, up substantially from $119 billion last quarter, as the chipmaker looks to secure memory and manufacturing capacity to meet AI demand over the next few years. That commitment includes $92 billion in projected spending for the rest of the fiscal year and another $87 billion in fiscal year 2028.
“The thing that matters for the industry is that AI is now doing productive and useful work,” Huang said during Wednesday’s call. “AI is generating profitable tokens…If we had more compute, we could generate more profitable tokens, which results in more profit for all of the services. This is the exact phase where we’re at, which is the reason why everybody’s leaning in.”
Investors will be watching to see if additional compute indeed translates so neatly into additional profits as AI companies pour hundreds of billions of dollars into infrastructure.